Skip to main content
Acta Pharmaceutica Sinica. B logoLink to Acta Pharmaceutica Sinica. B
. 2022 Nov 11;13(6):2572–2584. doi: 10.1016/j.apsb.2022.11.010

MF-SuP-pKa: Multi-fidelity modeling with subgraph pooling mechanism for pKa prediction

Jialu Wu a,b, Yue Wan c, Zhenxing Wu a,b, Shengyu Zhang c, Dongsheng Cao d,∗, Chang-Yu Hsieh a,∗, Tingjun Hou a,∗
PMCID: PMC10326308  PMID: 37425064

Abstract

Acid-base dissociation constant (pKa) is a key physicochemical parameter in chemical science, especially in organic synthesis and drug discovery. Current methodologies for pKa prediction still suffer from limited applicability domain and lack of chemical insight. Here we present MF-SuP-pKa (multi-fidelity modeling with subgraph pooling for pKa prediction), a novel pKa prediction model that utilizes subgraph pooling, multi-fidelity learning and data augmentation. In our model, a knowledge-aware subgraph pooling strategy was designed to capture the local and global environments around the ionization sites for micro-pKa prediction. To overcome the scarcity of accurate pKa data, low-fidelity data (computational pKa) was used to fit the high-fidelity data (experimental pKa) through transfer learning. The final MF-SuP-pKa model was constructed by pre-training on the augmented ChEMBL data set and fine-tuning on the DataWarrior data set. Extensive evaluation on the DataWarrior data set and three benchmark data sets shows that MF-SuP-pKa achieves superior performances to the state-of-the-art pKa prediction models while requires much less high-fidelity training data. Compared with Attentive FP, MF-SuP-pKa achieves 23.83% and 20.12% improvement in terms of mean absolute error (MAE) on the acidic and basic sets, respectively.

Key words: pKa prediction, Graph neural network, Subgraph pooling, Multi-fidelity learning, Data augmentation

Graphical abstract

Integrating domain knowledge into deep learning techniques significantly boosts model performance on pKa prediction. The proposed MF-SuP-pKa achieves state-of-the-art accuracy with higher robustness, and is capable of handling ionized molecules.

Image 1

1. Introduction

Acid-base dissociation constant (pKa), a fundamental physicochemical parameter describing the ionization ability of molecules, plays a vital role in chemical, environmental, agricultural, and medical sciences1,2. In drug discovery, pKa determines the ionization states of drug-like molecules under physiological environment, thus affecting a wide range of biopharmaceutical properties, such as lipophilicity, solubility, ADMET (absorption, distribution, metabolism, excretion, and toxicity), and target protein binding3, 4, 5. Considering that wet-lab experiments are costly and time-consuming, it's of great importance to develop rapid and accurate pKa prediction models.

It is well known that accurate prediction of pKa remains a formidable challenge. On the one hand, an amphoteric molecule could act as either an acid (proton donor) or a base (proton acceptor), and thus we need to build different quantitative structure–property relationship (QSPR) models for acidic pKa and basic pKa, separately. On the other hand, a multiprotic compound with n ionization sites will possess n macro-pKa and n × 2n−1 micro-pKa3. Micro-pKa measures the protonation or de-protonation transition of a specific titratable group, while macro-pKa reflects the ionization ability of the whole molecule. Fortunately, there exists a mathematical relationship between micro-pKa and macro-pKa [Eqs. (1), (2)]4

Acid:pKamacro=−Log(∑i=1N10−pKa(i)micro) (1)
Base:pKamacro=Log(∑i=1N10pKa(i)micro) (2)

In recent years, machine learning (ML) algorithms have been used to develop handcrafted feature-based pKa prediction models, such as eXtreme Gradient Boosting (XGBoost)1, Random Forest (RF)5, Deep Neural Networks (DNN)1,6, Support Vector Machine (SVM)7, and Radial Basis Function (RBF)8. These traditional ML models have two major deficiencies: (1) tedious feature engineering is often required, which is prone to introduce expert bias; (2) they can only handle fixed-size inputs and cannot fully utilize the internal structure information of compounds. Graph Neural Networks (GNN) are a set of deep learning (DL) algorithms that can learn representation automatically from graph-structured data for node-, subgraph- or graph-level tasks, and have achieved tremendous success in molecular property predictions9.

Due to the complexity of pKa, the applications of GNN in pKa predictions are relatively rare but promising. In 2019, Roszak et al.10 first proposed a GNN-based pKa predictor for the C–H bond, and the results demonstrate that GNN is useful in predicting atom-specific characteristics. In 2021, Pan et al.11 developed a pKa prediction web server called MolGpKa by combining GNN and a SMARTS list of ionizable substructures, aiming at identifying and predicting the pKa values for each ionizable group. However, MolGpKa was trained based on the computational pKa values in the ChEMBL database12, and the authors assigned the predicted pKa to the most acidic or most basic site using Epik13, which will inevitably introduce data noise when the major site selection was wrong or there was no dominant form for a multiprotic molecule. More recently, Xiong et al.4 proposed the Graph-pKa model based on the multi-instance learning (MIL) framework. The key point of Graph-pKa is predicting the micro-pKa from the node embeddings extracted from the GNN layers and then calculating the macro-pKa by the quantitative formula to compare with the ground truth. However, the ionization center identified by Graph-pKa is only limited to X–H (X is not carbon) acids and nitrogenous bases.

The above studies have verified the feasibility and superiority of GNN in pKa prediction, but they still have some shortcomings. First, both MolGpKa and Graph-pKa are restricted to predict the primary ionization ability of neutral molecules. However, the pKa values of ionized molecules are important in practice, especially for drug development. For instance, the antiallergic drug cetirizine with two basic sites and one acidic site will experience multi-step ionization under physiological conditions14. Second, the prediction of micro-pKa values is considered as a node-level regression task, which will break the inherent semantic information of ionizable groups (such as carboxyl groups)15 and neglect the long-range interaction at the intramolecular level16. Therefore, it's necessary to develop new graph-based methods that take advantage of chemical knowledge to achieve broader application and higher generalization ability.

In this study, to tackle the abovementioned challenges, we followed the MIL framework and proposed MF-SuP-pKa (multi-fidelity modeling with subgraph pooling for pKa prediction), a novel and effective model for pKa prediction. We extended the SMARTS list provided by Pan et al.11 to deal with ionized molecules and introduced a knowledge-aware subgraph pooling17 strategy to capture the graph context for micro-pKa prediction. These pre-defined ionization sites can serve as prior knowledge to boost model performance. Motivated by multi-fidelity modeling18,19, we incorporated low-fidelity data (computational pKa) into high-fidelity data (experimental pKa) through pre-training and fine-tuning to overcome the scarcity of accurate pKa data. Furthermore, according to the consistency between acidic pKa and basic pKa, a knowledge-guided data augmentation on the pre-training data set was implemented. The ablation study proved the effectiveness of subgraph pooling (SuP), multi-fidelity learning (MFL), and data augmentation (DA). Extensive evaluation on the internal and external test sets proved that MF-SuP-pKa achieved competitive performance for both the micro- and macro-pKa predictions. The visualization of the subgraph embeddings verified that MF-SuP-pKa is powerful enough to extract structure information related to the pKa values.

The main contributions of this work are as follows: (1) we formulate the micro-pKa prediction as a subgraph-level problem and integrate a powerful subgraph pooling strategy into the original Attentive FP model; (2) we alleviate data scarcity problem through multi-fidelity learning and data augmentation, and the proposed MF-SuP-pKa by using much less experimental data outperforms the state-of-the-art pKa prediction models; and (3) MF-SuP-pKa can provide more comprehensive prediction results, and the capacity of predicting micro-pKa and processing charged molecules makes multi-step pKa prediction feasible.

2. Materials and methods

2.1. Data sets

2.1.1. The extended SMARTS list of ionizable groups

The original SMARTS list provided by Pan et al.11 contains 143 titratable substructures in the neutral form. However, for ionized molecules, the protonated basic sites (considered as acidic sites) and the de-protonated acidic sites (considered as basic sites) could be the dominant ionization sites (Fig. 1A). To label the ionization centers more reasonably, we defined the nitrogen, oxygen, and sulfur atoms with positive formal charges and connected with at least one hydrogen atom as the acidic sites, and the nitrogen, oxygen, sulfur, and carbon atoms with negative formal charges as the basic sites (the oxygen atoms in nitro-group were excluded). These newly defined ionizable sites also provide conditions for data augmentation (Section 2.1.4). Furthermore, two common acidic groups in the DataWarrior data set (Section 2.1.3) were added (Fig. 1B). The final SMARTS list contains 148 titratable substructures.

Figure 1.

Figure 1

Extension of the SMARTS list of ionizable groups. (A) Illustration of the ionized sites acting as the dominant sites. Left: the protonated acidic site; Right: the de-protonated basic site. The black arrows indicate dissociation reactions. (B) The newly added acidic patterns that commonly appear in the DataWarrior data set.

2.1.2. Low-fidelity data for pre-training (the ChEMBL data set)

Our model was pre-trained based on the computational pKa values collected from the ChEMBL database (version 29)12, which are macro-pKa calculated using the ChemAxon Marvin Suite (version 17.29.0)20. ChEMBL contains around 2.1 million bioactive compounds, and the following criteria were used to select the pre-training data: (1) the type of compounds must be ‘small molecules’ and labeled as organic; (2) the molecular weight is between 100 and 799; (3) at least have a record of the most acidic or most basic pKa value; (4) the compounds containing multiple components (mixtures, salts, or solvents) were excluded.

The filtered ChEMBL data set consists of 1,535,537 molecules with 1,023,314 acidic pKa values and 1,109,725 basic pKa values. To reduce training time without sacrificing performance, we filtered the multiprotic molecules according to the ratio in the downstream tasks. In detail, for the acidic subset, all the monoprotic molecules (only the acidic sites were considered) were retained while the multiprotic molecules were randomly selected with a ratio of 2:1. Similarly, the monoprotic-multiprotic ratio was 1:1 for the basic subset. The final data set involves approximately 626.8 thousand acidic pKa values and 489.5 thousand basic pKa values. The ChEMBL data set was randomly split into the training and validation sets with a ratio of 8:2.

2.1.3. High-fidelity data for fine-tuning (the DataWarrior data set)

The high-fidelity experimental pKa data we used is the DataWarrior data set curated by Mansouri et al.7, and the data quality and structural diversity were validated in their study. To decrease data noise, the molecules with multiple distinct pKa values were removed. Based on the SMARTS list reported by Pan et al.11, those chemical structures without detectable ionization sites were removed. This yielded 5441 unique QSPR-ready structures, with 2750 acidic pKa values and 2992 basic pKa values. The DataWarrior data set was split into the training, validation, and test sets with a ratio of 8:1:1 for 10 times with different random seeds, and the models were trained and evaluated separately.

We clustered the amphoteric molecules (having both acidic and basic pKa) and selected 20 representative chemical structures to visualize the ionizable atom labels (Supporting Information Fig. S1). The distributions of the numbers of the ionization sites and the pKa values in the DataWarrior data set are illustrated in Fig. 2. The molecules could be divided into subtypes according to different criteria: one is the number of the ionization sites (monoprotic or multiprotic), and the other is the ionic states (neutral form or ionized form). The detailed information is reported in Table 1.

Figure 2.

Figure 2

The characteristics of the DataWarrior data set. (A) Distribution of the numbers of the ionization sites in the acidic set. (B) Distribution of the numbers of the ionization sites in the basic set. (C) Distribution of the experimental pKa values in the acidic and basic sets, respectively. Some amphoteric molecules with both acidic and basic pKa values are added into the two sets.

Table 1.

The detailed composition of the DataWarrior data seta.

Data set Monoprotic
Multiprotic
Neutral Ionized Neutral Ionized
Acidic 1812 30 558 350
Basic 1642 23 905 422
a

The ionized form includes zwitterion.

2.1.4. Knowledge-guided data augmentation

As defined above, we referred to the protonated basic sites as acidic sites and the de-protonated acidic sites as basic sites. Theoretically, the acidic pKa of the conjugate acid is equivalent to the basic pKa of the conjugate base. Therefore, we converted those monoprotic acids into their conjugate base forms, and vice versa (as illustrated in Fig. 3). This strategy resulted in the addition of 67,095 negatively-charged molecules to the ChEMBL basic set and 85,574 positively-charged molecules to the ChEMBL acidic set. Through data augmentation, we enriched the chemical types in the low-fidelity data set, which should offer help to the pKa prediction of ionized molecules in the downstream tasks.

Figure 3.

Figure 3

Illustration of the knowledge-guided data augmentation. The basic pKa of the conjugate base equals to the acidic pKa of the conjugate acid. Neutral molecules in the basic (acidic) set were turned to their ionized forms and added into the acidic (basic) set.

2.2. Overview of MF-SuP-pKa

Fig. 4 presents an overview of the MF-SuP-pKa model. We first summarize the multi-fidelity learning of model construction, and then we expand on the model architecture including input, graph encoding, subgraph-based readout and pKa prediction. In our model, we used two stacks of graph convolution layers to extract molecular representation: one is for node-level message passing, and the other is for subgraph-level message passing.

Figure 4.

Figure 4

Overview of MF-SuP-pKa. (A) The MF-SuP-pKa model was constructed using multi-fidelity learning: pre-trained on the ChEMBL data set and fine-tuned on the DataWarrior data set. (B) The prediction process consists of three steps: molecular graph construction and node-level convolution, subgraph pooling-based readout, and micro- and macro-pKa predictions. The updated embedding vector of each subgraph-level virtual node is the input to the fully connected neural networks for micro-pKa prediction, and macro-pKa is calculated according to the theoretical formula.

Let G=(V,Ԑ) denotes a molecular graph with a node set V(|V|=n) and an edge set Ԑ. The ionization site Ii is identified through substructure matching using the extended SMARTS list. For a target node Ii, its k-hop subgraph is defined as Si={v|d(Ii,v)≤k} where d(Ii,v) is the shortest path distance between Ii and node v. Given a molecule with m ionizable sites, the set of ionization centers is denoted by I=I1,I2,…Im, and the subgraph set is S=S1,S2,…Sm.

2.2.1. Multi-fidelity learning

As shown in Fig. 4A, the MF-SuP-pKa model was constructed by two phases: pre-training on the ChEMBL data set and fine-tuning on the DataWarrior data set. The target of the pre-training is to learn the QSPR between various chemical structures and the predicted pKa values, and the fine-tuning stage is aimed to migrate the predicted values to the experimental values. Distinct from self-supervised GNN pre-training21,22, the computational pKa and the experimental pKa are strongly correlated, thus avoiding the potential of negative transfer.

2.2.2. Input

The input to MF-SuP-pKa consists of four parts: (1) the node feature matrix X∈Rn×dnode containing nine types of atom features (Supporting Information Table S1); (2) the edge feature matrix E∈Rn×n×dedge containing four types of bond features (Table S1); (3) the adjacency matrix A∈{0,1}n×n; and (4) the topological distance matrix D∈Rn×n with Dij=d(vi,vj) measuring the length of the shortest path between vi and vj23.

2.2.3. GNN-based graph encoding

The graph convolution layer in MF-SuP-pKa is identical to Attentive FP16, owing to its brilliant performance in many molecular property prediction tasks, including pKa. In our pre-experiment, we evaluated three handcrafted feature-based methods, including RF24, XGBoost25, and DNN26, and three graph-based methods, including Graph Convolutional Networks (GCN)27, Message Passing Neural Networks (MPNN)28, and Attentive FP. On the DataWarrior data set, we found that Attentive FP achieved the best performance (see Supporting Information Table S2). In addition, the same convolution layer used in Graph-pKa also demonstrated the strength of Attentive FP.

Here, we use Attentiven (with 2 layers) for node-level message passing, and get the updated node embeddings as Eq. (3):

H=Attentiven(X,E,A) (3)

where H∈Rn×200 is the stack of hidden node features hv.

2.2.4. Subgraph-based readout

The key point of MF-SuP-pKa lies in the novel design of the readout phase, which consists of knowledge-aware subgraph pooling and subgraph-level message passing. In Graph-pKa, those updated node features are directly fed into the fully connected neural networks (FCNN) to predict the atom-level micro-pKa. Herein, we proposed a knowledge-aware subgraph pooling operation combined with the subgraph-level message passing to capture the local and global environments around the target node. Concretely, a virtual node is created connected with every atom in the molecule and its embedding is updated using a graph attention mechanism. There are three important differences compared to Attentive FP: (1) the number of the virtual nodes is equal to that of the ionization centers; (2) each virtual node denotes to a subgraph rather than the whole graph; and (3) to make the topologically closer atom contributes more to the target, the initial feature of each virtual node is the weighted sum of the atom features within the subgraph. The weight of each node is calculated by an exponential function using its shortest distance to the target ionization center. The extraction of the k-hop subgraph is centered on the pre-labeled ionization sites, and the optimal value of k was determined through hyperparameter search.

Firstly, the subgraph pooling operation is formulated as Eq. (4):

hSi=∑v∈Sihv×e−α×d(Ii,v),i=1,2,…m (4)

where hSi is the initial hidden state of the virtual node representing Si, α is a constant set to 0.1, and d(Ii,v) denotes to the shortest distance between the ionization center Ii and the node v in the subgraph Si.

Secondly, we apply subgraph-level message passing on the augmented graph with virtual nodes to fully extract the intrinsic and useful graph information. In the augmented graph, each virtual node corresponds to a k-hop subgraph, and the virtual edges linking the virtual nodes with all nodes in the original graph. Thus, the subgraph representations can be updated utilizing the message passing from the global environment. Specifically, we concatenate the node embeddings with the subgraph embeddings to get H′∈R(n+m)×200, and generate a new adjacency matrix A′∈{0,1}(n+m)×(n+m) representing the directed edges from the original graph nodes to the virtual nodes as in Eqs. (5), (6):

H′=[HHS] (5)
A′=[On×nOn×m1m×nOm×m] (6)

where HS∈Rm×200 denotes the stack of hSi, On×n, On×m, and Om×m are zero matrices, while 1m×n is an all-ones matrix.

Subsequently, AttentiveS (with 2 layers) is employed to get the updated subgraph embeddings Hs′∈Rm×200 as Eq. (7):

HS′=AttentiveS(H′,A′) (7)

2.2.5. Micro- and macro-pKa predictions

The updated subgraph-level embedding of each virtual node is the input for the pKa prediction module. Firstly, the micro-pKa is predicted by the one-layer FCNN as Eq. (8):

pKaSimicro=FCNNhSi, (8)

Secondly, the macro-pKa is calculated according to the micro-equilibria theory [Eqs. (1), (2)] as Eq. (9):

yˆ=pKamacro=δj∑i=1m10δj∙pKaSimicro (9)

where hSi, is the i-th row of HS′, and δj={−1,1} for acids and bases, respectively.

Finally, the predicted macro-pKa yˆ={yˆ1,yˆ2,…yˆN} is compared with the ground truth y={y1,y2,…yN} to calculate the mean square error loss as Eq. (10):

L=1N∑i=1N(yi−yˆi)2 (10)

2.3. Experimental settings

2.3.1. Baselines

We compared the performances of MF-SuP-pKa against Attentive FP and two state-of-the-art GNN-based pKa prediction tools. For the evaluation on the internal test set, Attentive FP and Graph-pKa were trained using the same DataWarrior data set. The pKa prediction from MolGpKa was obtained through the web server (https://xundrug.cn/molgpka) provided by Pan et al.11 Considering that ionized molecules are out of the applicability domains of MolGpKa and Graph-pKa, we reported the model performances on the entire test set and the neutral part, respectively.

2.3.2. Evaluation

The metrics for evaluating the model performance on macro-pKa prediction are coefficient of determination (R2), mean absolute error (MAE), and root mean squared error (RMSE). For micro-pKa prediction, we compared the dominant ionization sites predicted by MF-SuP-pKa with the reference sites assigned by Hunt et al.8 and reported the number of molecules predicted correctly.

2.3.3. Implementation

We utilized RDKit29 to extract the molecular features as the input for all models. The construction of the molecular graphs, MF-SuP-pKa model and baseline models were implemented by the DGL package with the PyTorch backend. The MF-SuP-pKa model was pre-trained on the ChEMBL data set for 100 epochs with a batch size of 512 and fine-tuned on the DataWarrior data set for 500 epochs with a batch size of 32. The hyperparameters of Attentive FP and Graph-pKa were set as same as those in the fine-tuning phase of our model, except that the number of graph neural layers was adjusted to six in Graph-pKa4. For parameter updating, we used Adam optimizer with a learning rate of 0.0003. The search range of hyperparameter k (the radius for subgraph extraction) was set to [0,1,2,3,4], and the optimal values were 2 and 3 for the acidic and basic models, respectively, as determined by 3 independent runs on the DataWarrior data set. To avoid overfitting, dropout was set to 0.1, and early-stopping was employed to maximize R2 on the validation set. Supporting Information Table S3 summarizes the hyperparameters in detail.

3. Results and discussion

This section provides a detailed analysis of the performance of MF-SuP-pKa and we try to answer the following five questions: (1) how does MF-SuP-pKa perform on macro-pKa prediction? (2) how does MF-SuP-pKa perform on micro-pKa prediction? (3) how do the subgraph pooling, multi-fidelity learning, and data augmentation influence the performance of MF-SuP-pKa? (4) whether the optimal value of hyperparameter k is reasonable and provides insightful interpretations? and (5) what kind of subgraph embeddings does MF-SuP-pKa learn?

3.1. Performance on macro-pKa prediction (Q1)

3.1.1. Internal test set

Table 2 summarizes the predictive performances of our proposed MF-SuP-pKa model and the baseline models, including Attentive FP, Graph-pKa, and MolGpKa.

Table 2.

Macro-pKa prediction performances on the internal test setsa.

Tasks Model Test set (all)
Test set (neutral)
R2 MAE RMSE R2 MAE RMSE
Acidic Attentive FP 0.885 ± 0.021 0.751 ± 0.055 1.141 ± 0.104 0.884 ± 0.027 0.741 ± 0.071 1.133 ± 0.129
Graph-pKa 0.599 ± 0.093 1.085 ± 0.111 2.124 ± 0.222 0.883 ± 0.030 0.707 ± 0.106 1.139 ± 0.165
MolGpKa 0.607 ± 0.063 1.151 ± 0.082 2.109 ± 0.151 0.716 ± 0.038 0.980 ± 0.076 1.779 ± 0.133
MF-SuP-pKa 0.919 ± 0.014 0.572 ± 0.020 0.956 ± 0.058 0.921 ± 0.015 0.553 ± 0.031 0.936 ± 0.077
Basic Attentive FP 0.884 ± 0.027 0.671 ± 0.067 1.063 ± 0.127 0.875 ± 0.028 0.679 ± 0.062 1.089 ± 0.120
Graph-pKa 0.601 ± 0.077 1.147 ± 0.083 1.974 ± 0.194 0.834 ± 0.032 0.798 ± 0.055 1.262 ± 0.145
MolGpKa 0.558 ± 0.051 1.287 ± 0.059 2.085 ± 0.112 0.671 ± 0.045 1.096 ± 0.065 1.781 ± 0.150
MF-SuP-pKa 0.916 ± 0.012 0.536 ± 0.028 0.910 ± 0.060 0.910 ± 0.013 0.534 ± 0.029 0.927 ± 0.059
a

Mean and standard deviation of the 10 independent runs are reported. The best results are depicted in bold.

For the acidic groups, the MF-SuP-pKa model achieved R2 of 0.919, MAE of 0.572, and RMSE of 0.956 on the complete test set, and it performed slightly better on the neutral test set (MAE = 0.553, RMSE = 0.936). For the basic groups, the MF-SuP-pKa model performed slightly worse on the entire test set (MAE = 0.536, RMSE = 0.910) than the neutral part (MAE = 0.534, RMSE = 0.927). Within the applicability domain of Graph-pKa and MolGpKa, these two models still obviously fell behind MF-SuP-pKa. Considering the ionized molecules, Graph-pKa and MolGpKa need to neutralize charges and then predict pKa values, and thus their performance on the entire test set dropped considerably.

The results revealed that our MF-SuP-pKa model showed great advantages on the internal test set, not only expanding the applicability domain from neutral molecules to ionized molecules but also outperforming the baselines on the neutral part. In addition, the standard deviation of the 10 independent runs of MF-SuP-pKa decreased significantly, demonstrating that our model achieved higher robustness.

Apart from the overall performance, the error distribution is also of great importance. Fig. 5A shows the ratio of the pKa values that are predicted within the absolute error margin in the test set. MF-SuP-pKa showed great advantages at a variety of thresholds, especially in lower error margins. In our acidic and basic models, 84% and 85% of the pKa values were predicted within ±1.0 log unit.

Figure 5.

Figure 5

Analysis of the error distribution on the DataWarrior test set. (A) The ratio of molecules predicted within the specific error margin in the prediction results of Attentive FP and MF-SuP-pKa. The ratio was averaged by 10 independent runs with different random seeds for data splitting and model initialization. (B) Distribution of the absolute errors (AE) predicted by MF-SuP-pKa with respect to the experimental acidic pKa. (C) Distribution of the AE predicted by MF-SuP-pKa with respect to the experimental basic pKa.

In-depth, the distributions of the absolute errors for acidic pKa and basic pKa are displayed in Fig. 5B and C, respectively. For the acidic model, MF-SuP-pKa performed best in the pKa ranges of 3–5 and 9–11, whereas the optimal ranges for the basic model were 4–6 and 8–11, which are highly consistent with the pKa distribution of the DataWarrior data set, implying that the model will provide more reliable predictions for the relevant pKa ranges.

3.1.2. External test set

Our model was further evaluated on three external data sets, namely the SAMPL6 data set, the SAMPL7 data set, and the Novartis data set.

The SAMPL630 and SAMPL731 data sets were established by Drug Design Data Resource Community and have been widely used for benchmarking pKa prediction methods32. The SAMPL6 data set contains 24 kinase inhibitor-like compounds with 12 acidic pKa and 19 basic pKa. The SAMPL7 data set consists of 22 molecules with 20 acidic pKa, most of which are sulfonamides. For a fair comparison, SM11 (with one basic pKa) and SM22 (with one acidic pKa and one basic pKa) were removed due to their presence in the DataWarrior data set. When it comes to Graph-pKa and MolGpKa, two pKa values (SM14_pKa1 and SM18_pKa3) not belonging to the neutral forms were further excluded. To estimate the ability of MF-SuP-pKa to handle ionized molecules, we compared it with the Hunt's model, which combines semi-empirical quantum mechanics (QM) descriptors and the RBF method for pKa prediction8. As illustrated in Fig. 6, the pKa values and the most acidic (basic) site of the neutral molecule were predicted first, and then the primary site was de-protonated (protonated) for secondary pKa prediction.

Figure 6.

Figure 6

The prediction processes of MF-SuP-pKa for SM14 and SM18. The circles denote to the ionization centers identified by the model. The red and blue numbers refer to the acidic and basic pKa values, respectively. In multi-step pKa prediction, the dominant ionization center (the basic center with the highest pKa value or the acidic center with the lowest pKa value) will be protonated or de-protonated for next prediction.

The Novartis data set provided by Baltruschat et al.5 comprises 280 pharmaceutical molecules with diverse chemical structures. The performance of MolGpKa on this data set is comparable to those of the commercial software Epik and Marvin11. According to the data source of Graph-pKa, the Novartis data set is part of its training data (the S-pKa data set, which is not open-source)4, and thus the evaluation was not conducted. As shown in Supporting Information Fig. S2, the pairwise Tanimoto similarities based on the ECFP4 fingerprints between the external test sets and the DataWarrior data set are lower than 0.4, suggesting that accurate prediction on the external test sets is a challenging task.

The prediction from the 10 independent runs of MF-SuP-pKa is listed in Supporting Information Table S4. Here, we selected a single model that performed well on both the SAMPL and Novartis data sets for a comparison. As shown in Table 3, MF-SuP-pKa exceeded or showed comparable accuracy with the counterparts on all the external test sets, highlighting the effectiveness of our model in real-world scenarios. Notably, MF-SuP-pKa showed relatively poor MAE on the SAMPL6 data set, which can be contributed from the distribution of deviations. As shown in Fig. 7, except SM21 with the largest deviation (−1.38), the differences between the predictions and experimental data for the other molecules are within or slightly higher than ±1 log unit, which is also consistent with the lower RMSE and higher R2 of MF-SuP-pKa compared with the other models.

Table 3.

Macro-pKa prediction performance on the external test setsa.

Data sets Model R2 MAE RMSE
SAMPL6 (26 pKa) MolGpKa(b) 0.907 0.522 0.773
Graph-pKa(c) 0.912 0.625 0.753
MF-SuP-pKa 0.912 0.687 0.751
SAMPL6 (28 pKa) Hunt's model(d) 0.894 0.698 0.880
MF-SuP-pKa 0.921 0.694 0.758
SAMPL7 (20 pKa) MolGpKa(b) 0.826 0.797 0.980
Graph-pKa(c) 0.839 0.758 0.934
MF-SuP-pKa 0.879 0.656 0.816
Novartis Acidic (112 pKa) MolGpKa(e) 0.75 0.85 1.29
MF-SuP-pKa 0.82 0.85 1.09
Novartis Basic (168 pKa) MolGpKa(e) 0.75 0.79 1.06
MF-SuP-pKa 0.86 0.61 0.79
a

The best results are highlighted in bold. Data source.

b

The web server of MolGpKa (https://xundrug.cn/molgpka/).

c

The predicted values provided in the original article of Graph-pKa7, the final metrics were averaged among three independent runs.

d

The predicted values are provided in the original article of Hunt's model3.

e

Cited from the article of MolGpKa13.

Figure 7.

Figure 7

Plot of the computational pKa values predicted by MF-SuP-pKaversus experimental pKa values for the SAMPL6 data set. Dotted lines show the errors that are within ±1 log unit. SM21 is the compound with the largest absolute error.

The absolute error distribution of the MF-SuP-pKa model on the Novartis data set is listed in Table 4. Regarding 1 log unit of AE as the threshold for acceptable errors, 60.7% of the Novartis acidic set and 76.8% of the Novartis basic set are within the applicability domain of our model. In the Novartis acidic set, half of the outliers (AE>2 log unit) were overestimated (N01–N05 in Fig. 8) while the others were underestimated (N06–N10 in Fig. 8). The conjugated effect in N01–N03 and the intra-molecular hydrogen bonds in N04–N05 will stabilize the anions and strongly enhance the acidity, thus making the prediction more difficult. As for N06–N08 with the dione chemotype and N09 with the hydroxyurea group, the large deviations are possibly caused by the lack of examples in the training set. Significant underestimate for N10 (−2.27) was observed, likely related to the vinylogous amide substructure inside the dashed box. According to the principle of Vinylogy, cations will be stabilized by electron delocation, thus making the amine a much stronger base and reducing the acidity of the molecule33. The outlier analysis can reflect the inadequacies of our model and guide further improvement.

Table 4.

Absolute error (AE) distribution on the Novartis data set.

Data sets Model Max AE AE distribution
<1 1–2 >2
Novartis acidic (112 pKa) MF-SuP-pKa 2.70 68 34 10
Novartis basic (168 pKa) MF-SuP-pKa 2.23 129 36 3
Figure 8.

Figure 8

The outliers (AE>2 log units) in the Novartis acidic set predicted by MF-SuP-pKa. N01–N05 were overestimated and N06–N10 were underestimated. The orange dashed boxes indicate the key substructures influencing the pKa values.

3.2. Performance on micro-pKa prediction (Q2)

Micro-pKa prediction is useful in multi-step pKa prediction, where we should adjust the protonation state of the query molecule and estimate the corresponding pKa iteratively. In addition, the accurate prediction of the dominant microstate (tautomer) is essential for modeling protein-ligand interactions34,35. However, the lack of experimental pKa data limits the construction and evaluation of micro-pKa prediction tools. The MIL framework for micro-pKa prediction in Graph-pKa is a great practice. In this part, we examined the micro-pKa prediction accuracy on the SAMPL6 data set by comparing the predicted dominant sites with the expert-labeled sites. We took the expert judgments from Hunt et al.8 as a reference and revised the labels of five benzimidazole and 4-aminoquinazoline derivatives (SM02, SM04, SM09, SM07, and SM14) according to the evidence of Nuclear Magnetic Resonance (NMR)32. Besides, the most basic site of SM17 was adjusted from 1,3,4-Oxadiazol ring to the pyridine ring based on the analysis of Xiong et al.4 Under this circumstance, the number of the consistent molecules versus the random split time is plotted in Fig. 9. Generally, MF-SuP-pKa provided acceptable results for distinguishing the strength of different ionization sites. In comparison with Graph-pKa, our model achieved better performance in 1 and 7 of the 10 times on the acidic and basic tasks, respectively. It's impressive that we only used a subset of the S-pKa data set (5742 versus 17,489 pKa values) to construct MF-SuP-pKa. The reference labels and micro-pKa prediction results are displayed in Supporting Information Fig. S3.

Figure 9.

Figure 9

Comparison of the model performance on micro-pKa prediction. The total number of the multiprotic molecules is 17 for the basic task and 4 for the acidic task. The consistency rates of Graph-pKa are 9/17 and 3/4 for the basic and acidic tasks, respectively. MF-SuP-pKa correctly predicted 7–13 dominant basic sites and 3–4 dominant acidic sites under different random split times.

3.3. Ablation study (Q3)

In this subsection, we conducted an ablation study to evaluate the contribution of different strategies in our model. We compared the MF-SuP-pKa model with its variants without data augmentation, multi-fidelity learning (named SuP-pKa), and subgraph pooling (the original Attentive FP model). As shown in Table 5, the three strategies are all effective to the model performance, indicating that the chemical environment extracted by subgraph pooling was superior to the original operation in Attentive FP, and the data scarcity problem was alleviated by multi-fidelity learning and data augmentation. Compared with the self-supervised pre-training (taking the GROVER model36 for example, see Supporting Information Table S5), the multi-fidelity learning successfully avoided the negative transfer.

Table 5.

Ablation results on the internal test set.

Task Strategy
Metric
Relative gain (MAE)
SuPa MFLb DAc R2 MAE RMSE
Acidic × × × 0.885 ± 0.021 0.751 ± 0.055 1.141 ± 0.104 /
√ × × 0.899 ± 0.024 0.669 ± 0.070 1.065 ± 0.116 10.92%
√ √ × 0.912 ± 0.021 0.589 ± 0.033 0.992 ± 0.104 21.57%
√ √ √ 0.919 ± 0.014 0.572 ± 0.020 0.956 ± 0.058 23.83%
Basic × × × 0.884 ± 0.027 0.671 ± 0.067 1.063 ± 0.127 /
√ × × 0.895 ± 0.018 0.638 ± 0.054 1.014 ± 0.095 4.92%
√ √ × 0.906 ± 0.021 0.554 ± 0.050 0.958 ± 0.108 17.44%
√ √ √ 0.916 ± 0.012 0.536 ± 0.028 0.910 ± 0.060 20.12%

Three strategies used in MF-SuP-pKa model.

a

SuP: subgraph pooling.

b

MFL: Multi-fidelity learning.

c

DA: data augmentation.

A more detailed analysis of the model performance on different kinds of molecules is shown in Fig. 10, verifying that our model outperforms Attentive FP on all subsets. The data augmentation strategy, in particular, aided us in achieving the initial goal of improving performance on ionized molecules. Furthermore, the pKa prediction of neutral compounds benefited as well. It's noteworthy that the basic model achieved less gain via subgraph pooling, mainly reflected by fewer improvements for the monoprotic molecules (Fig. 10B). One possible reason is that the molecular size is tended to be smaller for the monoprotic compounds (Supporting Information Fig. S4), implying that the virtual node extracted from the entire molecule or the 3-hop subgraph will be more similar.

Figure 10.

Figure 10

The performance of the MF-SuP-pKa model and its variants on different types (neutral versus ionized or monoprotic versus multiprotic) of molecules in (A) the DataWarrior acidic set and (B) the DataWarrior basic set.

3.4. Hyperparameter sensitivity (Q4)

We conducted hyperparameter search on the Attentive FP model augmented with the subgraph pooling mechanism (but w/o multi-fidelity learning and data augmentation) to optimize model performance. As shown in Fig. 11, the best values of k for the acidic and basic models were determined as 2 and 3, respectively. It's reasonable because the characteristics of acidic and basic functional groups are naturally distinct.

Figure 11.

Figure 11

Hyperparameter sensitivity of the Attentive FP model augmented with the subgraph pooling mechanism on (A) the MAE metric and (B) the R2 metric. The hyperparameter k refers to the radius for subgraph extraction.

For instance, the basic groups commonly appear in heterocycles (pyridine, pyrimidine, pyrazine, quinoline, etc.), and the 3-hop subgraph excels in describing the local chemical environment of those ionization centers. For the acidic pKa prediction, however, the 2-hop neighbors are sufficient to distinguish acidic groups in different graph contexts, like a carboxylic group connected with an aliphatic or an aromatic system (Fig. 12).

Figure 12.

Figure 12

Typical acidic and basic ionization groups. Left: N-containing basic functional groups that can be differentiated by a subgraph with a radius of 3; Right: O-containing acidic functional groups that can be differentiated by a subgraph with a radius of 2.

3.5. Visualization of the subgraph embeddings (Q5)

To figure out what the MF-SuP-pKa model learned in the pre-training and fine-tuning stages, we analyzed the subgraph embeddings generated by the pre-trained model and the final MF-SuP-pKa model separately. Specifically, we used t-distributed stochastic neighbor embedding (t-SNE)37 to visualize the updated subgraph embeddings (200-dimensional vectors) of the 1501 monoprotic molecules in the DataWarrior training set. As shown in Fig. 13, the pre-trained model can roughly distinguish the subgraphs of different pKa ranges, which is beneficial for the downstream tasks. However, we can see that some weak-acid points (in blue color) are mixed with strong-acid points (in orange color), which is possibly due to the deviation between the computational and experimental pKa values. After fine-tuning, the distribution of the subgraph embeddings is differentiated by the pKa values more clearly and ideally. The above results prove that our MF-SuP-pKa model can capture better representation and perform well in the subgraph-level regression task.

Figure 13.

Figure 13

The t-SNE plot of the subgraph embedding vectors of the 1501 monoprotic molecules in the DataWarrior training set learned by (A) the pre-trained model and (B) the MF-SuP-pKa model. The points are colored by experimental pKa values.

3.6. Data availability

The data sets used in this work, including ChEMBL data set, DataWarrior data set, and three external test sets (SAMPL 6 data set, SAMPL 7 data set, Novartis data set), are available at https://github.com/wujialu/MF-SuP-pKa. The original QSPR-ready DataWarrior data set provided by Mansouri et al. can be downloaded from https://github.com/NIEHS/OPERA.

3.7. Code availability

Weights for the pre-trained model and the source codes of MF-SuP-pKa are available in the GitHub repository: https://github.com/wujialu/MF-SuP-pKa.

4. Conclusions

In this work, we proposed MF-SuP-pKa, a domain knowledge-integrated GNN model for pKa prediction. Specifically, we identified the ionization centers explicitly using the extended SMARTS list, and implemented the knowledge-guided data augmentation and knowledge-aware subgraph pooling. In addition, we took full advantage of the low-fidelity data to improve the model's capacity to fit the high-fidelity data through transfer learning, which will not only facilitate the training process but also achieve high robustness and generalization ability. We systematically analyzed the model performance on macro-pKa and micro-pKa prediction, and the results showed that MF-SuP-pKa achieved excellent performance on the internal and external test sets. The ablation study confirmed the validity of the three strategies to incorporate chemical domain knowledge into the DL framework. Moreover, the analysis of hyperparameter sensitivity and the visualization of subgraph embeddings provided a chemical insight into the model performance. In contrast to Graph-pKa and MolGpKa, we preserved the protonation state of the molecules and treated the charged atom as a potential dissociation site, thus providing a viable option to deal with both neutral and ionized molecules and paving the way for multi-step prediction. As a final note, MF-SuP-pKa compared favorably with Graph-pKa using only one-third of the training data, we can expect that the model performance will be further improved as the quality and quantity of pKa data grow.

Acknowledgments

This work was financially supported by National Key Research and Development Program of China (2021YFF1201400), National Natural Science Foundation of China (22220102001), and Natural Science Foundation of Zhejiang Province (LZ19H300001, LD22H300001, China).

Footnotes

Peer review under the responsibility of Chinese Pharmaceutical Association and Institute of Materia Medica, Chinese Academy of Medical Sciences.

Appendix A

Supporting data to this article can be found online at https://doi.org/10.1016/j.apsb.2022.11.010.

Contributor Information

Dongsheng Cao, Email: oriental-cds@163.com.

Chang-Yu Hsieh, Email: kimhsieh@zju.edu.cn.

Tingjun Hou, Email: tingjunhou@zju.edu.cn.

Author contributions

Jialu Wu, Yue Wan and Zhenxing Wu designed and performed experiments, analyzed data, and wrote the paper. Shengyu Zhang performed some experiments. Dongsheng Cao, Chang-Yu Hsieh and Tingjun Hou initiated the study, organized, designed, and wrote the paper. All authors read and approved the final manuscript.

Conflicts of interest

The authors declare that they have no competing interests.

Appendix A. Supplementary data

The following is the Supplementary data to this article:

Multimedia component 1
mmc1.pdf (692.8KB, pdf)

References

  • 1.Yang Q., Li Y., Yang J.D., Liu Y., Zhang L., Luo S., et al. Holistic prediction of the pKa in diverse solvents based on a machine-learning approach. Angew Chem Int Ed. 2020;59:19282–19291. doi: 10.1002/anie.202008528. [DOI] [PubMed] [Google Scholar]
  • 2.Lu Y., Anand S., Shirley W., Gedeck P., Kelley B.P., Skolnik S., et al. Prediction of pKa using machine learning methods with rooted topological torsion fingerprints: application to aliphatic amines. J Chem Inf Model. 2019;59:4706–4719. doi: 10.1021/acs.jcim.9b00498. [DOI] [PubMed] [Google Scholar]
  • 3.Rupp M., Korner R., Tetko I V. Predicting the pKa of small molecules. Comb Chem High T Scr. 2011;14:307–327. doi: 10.2174/138620711795508403. [DOI] [PubMed] [Google Scholar]
  • 4.Xiong J., Li Z., Wang G., Fu Z., Zhong F., Xu T., et al. Multi-instance learning of graph neural networks for aqueous pKa prediction. Bioinformatics. 2022;38:792–798. doi: 10.1093/bioinformatics/btab714. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 5.Baltruschat M., Czodrowski P. Machine learning meets pKa. F1000Res. 2020;9 doi: 10.12688/f1000research.22090.1. Chem Inf Sci-113. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 6.Fraczkiewicz R., Lobell M., Göller A.H., Krenz U., Schoenneis R., Clark R.D., et al. Best of both worlds: combining pharma data and state of the art modeling technology to improve in silico pKa prediction. J Chem Inf Model. 2015;55:389–397. doi: 10.1021/ci500585w. [DOI] [PubMed] [Google Scholar]
  • 7.Mansouri K., Cariello N.F., Korotcov A., Tkachenko V., Grulke C.M., Sprankle C.S., et al. Open-source QSAR models for pKa prediction using multiple machine learning approaches. J Cheminformatics. 2019;11:60. doi: 10.1186/s13321-019-0384-1. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 8.Hunt P., Hosseini-Gerami L., Chrien T., Plante J., Ponting D.J., Segall M. Predicting pKa using a combination of semi-empirical quantum mechanics and radial basis function methods. J Chem Inf Model. 2020;60:2989–2997. doi: 10.1021/acs.jcim.0c00105. [DOI] [PubMed] [Google Scholar]
  • 9.Wieder O., Kohlbacher S., Kuenemann M., Garon A., Ducrot P., Seidel T., et al. A compact review of molecular property prediction with graph neural networks. Drug Discov Today Technol. 2020;37:1–12. doi: 10.1016/j.ddtec.2020.11.009. [DOI] [PubMed] [Google Scholar]
  • 10.Roszak R., Beker W., Molga K., Grzybowski B.A. Rapid and accurate prediction of pKa values of C–H acids using graph convolutional neural networks. J Am Chem Soc. 2019;141:17142–17149. doi: 10.1021/jacs.9b05895. [DOI] [PubMed] [Google Scholar]
  • 11.Pan X., Wang H., Li C., Zhang J.Z.H., Ji C. MolGpka: a web server for small molecule pKa prediction using a graph-convolutional neural network. J Chem Inf Model. 2021;61:3159–3165. doi: 10.1021/acs.jcim.1c00075. [DOI] [PubMed] [Google Scholar]
  • 12.Gaulton A., Bellis L.J., Bento A.P., Chambers J., Davies M., Hersey A., et al. ChEMBL: a large-scale bioactivity database for drug discovery. Nucleic Acids Res. 2012;40:1100–1107. doi: 10.1093/nar/gkr777. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 13.Shelley J.C., Cholleti A., Frye L.L., Greenwood J.R., Timlin M.R., Uchimaya M. Epik: a software program for pKa prediction and protonation state generation for drug-like molecules. J Comput Aid Mol Des. 2007;21:681–691. doi: 10.1007/s10822-007-9133-z. [DOI] [PubMed] [Google Scholar]
  • 14.Navo C.D., Jiménez-Osés G. Computer Prediction of pKa values in small molecules and proteins. ACS Med Chem Lett. 2021;12:1624–1628. doi: 10.1021/acsmedchemlett.1c00435. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 15.Zhang Z, Liu Q, Wang H, Lu C, Lee CK. Motif-based graph self-supervised learning for molecular property prediction. arXiv:2110.00987. Available from: 10.48550/arXiv.2110.00987. [DOI]
  • 16.Xiong Z., Wang D., Liu X., Zhong F., Wan X., Li X., et al. Pushing the boundaries of molecular representation for drug discovery with the graph attention mechanism. J Med Chem. 2020;63:8749–8760. doi: 10.1021/acs.jmedchem.9b00959. [DOI] [PubMed] [Google Scholar]
  • 17.Crouse M, Abdelaziz I, Cornelio C, Thost V, Wu L, Forbus K, et al. Improving graph neural network representations of logical formulae with subgraph pooling. arXiv:1911.06904. Available from: https://arxiv.org/abs/1911.06904.
  • 18.Hebbal A., Brevault L., Balesdent M., Talbi E.G., Melab N. Multi-fidelity modeling with different input domain definitions using deep Gaussian processes. Struct Multidiscip O. 2021;63:2267–2288. [Google Scholar]
  • 19.Xu Y., Song X., Zhang C. Hierarchical regression framework for multi-fidelity modeling. Knowl Base Syst. 2021;212 [Google Scholar]
  • 20.ChemAxon Marvin Suite, ChemAxon Inc. Available from: http://www.chemaxon.com.
  • 21.Hu W, Liu B, Gomes J, Zitnik M, Liang P, Pande V, et al. Strategies for pre-training graph neural networks. arXiv: 1905.12265. Available from: https://arxiv.org/abs/1905.12265..
  • 22.Li H., Zhao X., Li S., Wan F., Zhao D., Zeng J. MoTSE: an interpretable task similarity estimator for small molecular property prediction tasks. bioRxiv. 2021 doi: 10.1101/2021.01.13.426608. Available from: [DOI] [Google Scholar]
  • 23.Chen J, Zheng S, Song Y, Rao J, Yang Y. Learning attributed graph representations with communicative message passing transformer. arXiv: 2107.08773. Available from: https://arxiv.org/abs/2107.08773.
  • 24.Svetnik V., Liaw A., Tong C., Culberson J.C., Sheridan R.P., Feuston B.P. Random forest: a classification and regression tool for compound classification and QSAR modeling. J Chem Inf Comput Sci. 2003;43:1947–1958. doi: 10.1021/ci034160g. [DOI] [PubMed] [Google Scholar]
  • 25.Chen T., Guestrin C. Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining. 2016. XGBoost: a scalable tree boosting system; pp. 785–794. [Google Scholar]
  • 26.Ma J., Sheridan R.P., Liaw A., Dahl G.E., Svetnik V. Deep neural nets as a method for quantitative structure–activity relationships. J Chem Inf Model. 2015;55:263–274. doi: 10.1021/ci500747n. [DOI] [PubMed] [Google Scholar]
  • 27.Kipf TN, Welling M. Semi-supervised classification with graph convolutional networks. arXiv: 1609.02907. Available from: https://arxiv.org/abs/1609.02907
  • 28.Gilmer J, Schoenholz SS, Riley PF, Vinyals O, Dahl GE. Neural message passing for quantum chemistry. arXiv:1704.01212. Available from: 10.48550/arXiv.1704.01212. [DOI]
  • 29.Landrum G. RDKit: open-source cheminformatics. Available from: https://www.rdkit.org.
  • 30.Işık M., Levorse D., Rustenburg A.S., Ndukwe I.E., Wang H., Wang X., et al. pKa measurements for the SAMPL6 prediction challenge for a set of kinase inhibitor-like fragments. J Comput Aid Mol Des. 2018;32:1117–1138. doi: 10.1007/s10822-018-0168-0. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 31.Drug Design Data Resource Community. The SAMPL7 data set. Version 1.1. Available from: https://zenodo.org/record/5637494#.Y0AXD7ZBxsY.
  • 32.Işık M., Rustenburg A.S., Rizzi A., Gunner M.R., Mobley D.L., Chodera J.D. Overview of the SAMPL6 pKa challenge: evaluating small molecule microscopic and macroscopic pKa predictions. J Comput Aid Mol Des. 2021;35:131–166. doi: 10.1007/s10822-020-00362-6. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 33.Perrin D.D., Dempsey B., Serjeant E.P. Springer Netherlands; Dordrecht: 1981. pKa prediction for organic acids and bases. [Google Scholar]
  • 34.Ropp P.J., Kaminsky J.C., Yablonski S., Durrant J.D. Dimorphite-DL: an open-source program for enumerating the ionization states of drug-like small molecules. J Cheminformatics. 2019;11:14. doi: 10.1186/s13321-019-0336-9. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 35.Watson M.A., Yu H.S., Bochevarov A.D. Generation of tautomers using micro-pKa’s. J Chem Inf Model. 2019;59:2672–2689. doi: 10.1021/acs.jcim.8b00955. [DOI] [PubMed] [Google Scholar]
  • 36.Rong Y, Bian Y, Xu T, Xie W, Wei Y, Huang W, et al. Self-supervised graph transformer on large-scale molecular data. In proceedings of 34th Conference on Neural Information Processing Systems (NeurIPS 2020), Vancouver, Canada. Available from: https://proceedings.neurips.cc/paper/2020/hash/94aef38441efa3380a3bed3faf1f9d5d-Abstract.html
  • 37.Van der Maaten L., Hinton G. Visualizing data using t-SNE. J Mach Learn Res. 2008;9:2579–2605. [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Multimedia component 1
mmc1.pdf (692.8KB, pdf)

Data Availability Statement

The data sets used in this work, including ChEMBL data set, DataWarrior data set, and three external test sets (SAMPL 6 data set, SAMPL 7 data set, Novartis data set), are available at https://github.com/wujialu/MF-SuP-pKa. The original QSPR-ready DataWarrior data set provided by Mansouri et al. can be downloaded from https://github.com/NIEHS/OPERA.


Articles from Acta Pharmaceutica Sinica. B are provided here courtesy of Elsevier

RESOURCES